Experience in preparing Web content for people with visual disabilities shows that automatic testing of accessibility is clearly insufficient to meet real-world demands, and that human reviews are needed. At the same time automated or semi-automated testing is an important part of reducing the demands on an human evaluator. In this paper we discuss tools we have developed with a design approach that addresses these points. As a consequence, we propose a flexible tool architecture, based on the integration of evaluation and repair functions. KEYWORDS Accessibility, architecture model, human review, automatic evaluation, EARL/RDF